Function Calling

Function calling lets a model ask your application to run code. You describe each function with a name, a description and a JSON Schema for its parameters. When the model decides a function is needed, it does not run anything itself. It returns the function name and a JSON arguments object, your code executes the call, and you send the result back so the model can continue or write its final answer. OpenAI and Google call this function calling, and Anthropic calls the same pattern tool use.

Two controls matter in production. Schema enforcement: OpenAI strict mode makes calls follow your schema instead of best effort, though Chat Completions stays non strict by default, and Anthropic offers strict: true on a tool definition. Tool choice: OpenAI tool_choice accepts auto, required, one forced function or an allowed_tools subset, and Gemini offers auto, any, none and validated modes. A model can also request several calls in one turn, which you can switch off.

import json


def get_order_status(order_id: str) -> dict:
    return {"order_id": order_id, "status": "shipped"}


TOOLS = {"get_order_status": get_order_status}

# Shape of a function call item from the OpenAI Responses API (mocked)
call = {
    "type": "function_call",
    "call_id": "call_123",
    "name": "get_order_status",
    "arguments": '{"order_id": "A1001"}',
}

result = TOOLS[call["name"]](**json.loads(call["arguments"]))
print({"type": "function_call_output", "call_id": call["call_id"], "output": json.dumps(result)})

The dispatch step of a function calling loop. The call is mocked. In a real loop it comes from the model response, and the output goes back in the next request.

Use function calling whenever the model must read live data or cause a side effect, such as an order lookup or a database query. Keep descriptions specific, because the model picks tools from them. Strict mode enforces the shape of the arguments, not that the values are allowed, so check permissions in your own code before any side effect. Skip it when the steps are fixed and known. Plain code is simpler.

• OpenAI: Function calling guide
• Anthropic: Tool use overview
• Google: Function calling with the Gemini API

Related Terms